fix(anthropic): carry the served model from message_start onto stream chunks

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

litellm/llms/anthropic/chat/handler.py

@@ -632,6 +632,7 @@ class ModelResponseIterator:
         self.tool_name_reverse_map: dict[str, str] = tool_name_reverse_map or {}
         # Generate response ID once per stream to match OpenAI-compatible behavior
         self.response_id = _generate_id()
+        self.served_model: str | None = None
 
         # Track if we're currently streaming a response_format tool
         self.is_response_format_tool: bool = False
@@ -1067,6 +1068,8 @@ class ModelResponseIterator:
                 }
                 """
                 message_start_block: Final = MessageStartBlock(**chunk)
+                start_message: Final = message_start_block["message"]
+                self.served_model = start_message["model"] if "model" in start_message else None
                 if "usage" in message_start_block["message"]:
                     usage = self._handle_usage(anthropic_usage_chunk=message_start_block["message"]["usage"])
             elif type_chunk == "error":
@@ -1098,6 +1101,7 @@ class ModelResponseIterator:
                 ],
                 usage=usage,
                 id=self.response_id,
+                model=self.served_model,
             )
 
             return returned_chunk

tests/test_litellm/llms/anthropic/chat/test_anthropic_chat_handler.py

@@ -2719,3 +2719,81 @@ class TestRustChatCompletionsHook:
             "model": "m",
             "messages": [],
         }
+
+
+def _served_model_stream_chunks(model: str | None) -> list[dict]:
+    message: Final = {
+        "id": "msg_served",
+        "type": "message",
+        "role": "assistant",
+        "content": [],
+        "usage": {"input_tokens": 10, "output_tokens": 1},
+    }
+    if model is not None:
+        message["model"] = model
+    return [
+        {"type": "message_start", "message": message},
+        {
+            "type": "content_block_start",
+            "index": 0,
+            "content_block": {"type": "text", "text": ""},
+        },
+        {
+            "type": "content_block_delta",
+            "index": 0,
+            "delta": {"type": "text_delta", "text": "Hello"},
+        },
+        {"type": "content_block_stop", "index": 0},
+        {
+            "type": "message_delta",
+            "delta": {"stop_reason": "end_turn"},
+            "usage": {"output_tokens": 2},
+        },
+        {"type": "message_stop"},
+    ]
+
+
+def test_message_start_model_is_carried_on_stream_chunks():
+    iterator: Final = ModelResponseIterator(None, sync_stream=True)
+
+    parsed: Final = [
+        iterator.chunk_parser(chunk) for chunk in _served_model_stream_chunks("claude-served-1")
+    ]
+
+    assert all(chunk.model == "claude-served-1" for chunk in parsed)
+
+
+def test_message_start_without_model_leaves_chunk_model_unset():
+    iterator: Final = ModelResponseIterator(None, sync_stream=True)
+
+    parsed: Final = [
+        iterator.chunk_parser(chunk) for chunk in _served_model_stream_chunks(None)
+    ]
+
+    assert all(chunk.model is None for chunk in parsed)
+
+
+def test_served_model_reaches_assembled_stream_through_custom_stream_wrapper():
+    from litellm.litellm_core_utils.streaming_handler import CustomStreamWrapper
+
+    served_model: Final = "claude-served-1"
+    sse_lines: Final = [
+        f"data: {json.dumps(chunk)}\n".encode() for chunk in _served_model_stream_chunks(served_model)
+    ]
+    iterator: Final = ModelResponseIterator(iter(sse_lines), sync_stream=True)
+    wrapper: Final = CustomStreamWrapper(
+        completion_stream=iter(iterator),
+        model="anthropic/claude-requested",
+        custom_llm_provider="anthropic",
+        logging_obj=MagicMock(),
+    )
+
+    chunks: Final = list(wrapper)
+
+    assert len(chunks) > 1
+    for chunk in chunks[1:]:
+        assert chunk._hidden_params["provider_response_model"] == served_model
+    assembled: Final = litellm.stream_chunk_builder(
+        chunks=list(chunks), messages=[{"role": "user", "content": "hi"}]
+    )
+    assert assembled._hidden_params["provider_response_model"] == served_model